Atlas

Benchmarks

← All benchmarks

Phare Encoding Jailbreaks

Safety · 2025-12-09

Phare jailbreak task using encoded or obfuscated prompts to bypass safeguards.

Top models (higher is better)

ModelScore
Sonnet 571.1
Llama 3.1 8B Instruct65.9
Magistral Small 1.264.4
Llama 3.1 405B Instruct63.0
Qwen3 8B61.4
Magistral Medium 1.258.1
Opus 4.151.8
GPT-4o44.4
Opus 4.543.6
Qwen3.7-Plus42.4
Gemini 3 Pro Preview42.1
Haiku 4.541.9
Qwen3.7 Max Preview41.9
Gemma 3 12B IT41.1
Gemma 4 31B IT38.2
Mistral Small 3.2 24B Instruct 250637.5
Llama 4 Scout37.4
Grok 436.8
Llama 3.3 70B Instruct33.7
Qwen3 VL 30B A3B Instruct33.6
Mistral Medium 3.533.4
Kimi K2.633.3
Opus 4.633.2
Grok 3 Mini33.1
Sonnet 4.532.1
Command A32.0
Gemma 3 27B IT32.0
GPT-4.1 Nano30.6
GPT-4o Mini29.8
GPT-529.5
Mistral Medium 3.128.3
GPT-5.227.8
Qwen2.5-Max27.6
Mistral Large 2 (Instruct 2407)27.6
gpt-oss-120b26.3
DeepSeek-V4-Flash25.3
GPT-5 Nano24.2
DeepSeek-R1-052824.0
Grok 224.0
Llama 4 Maverick23.5
Loading Atlas data…